Enhancing Indoor Human Localization with Echo State Networks and Temporal Convolutional Networks

Indoor human localization has gained significant importance due to its wide range of applications, including smart environments, health monitoring, and security systems. Accurate indoor positioning remains a challenging task due to signal interference, complex environments, and the need for real-time data processing. Localization systems can be classified into active and passive methods, each with unique characteristics and applications. Active systems require external signal emissions, using transmitters and receivers to determine an object or person’s location. While effective, these systems involve higher infrastructure costs due to additional equipment, and they require regular maintenance and user interaction, which can impact convenience. In contrast, passive localization systems rely on natural signals or environmental interactions, reducing installation and maintenance costs, and offering a less intrusive experience as no wearable tags are needed. Passive methods such as infrared, ultrasound, and radar sensing, though efficient, often face limitations like high costs or privacy concerns. Capacitive sensors, particularly those using long-range load-mode operation, present a promising alternative for indoor localization. These sensors are affordable, discreet, and preserve privacy by using the human body as a conductive element. However, they are vulnerable to environmental noise, affecting their accuracy and range. Enhancing noise resistance through improved sensor design and advanced signal processing can help overcome these limitations. For this thesis I investigate the possibility of using Echo State Network for localization with data acquired with this type of sensors. First, I studied the Echo State Networks (ESN) and the effects of their hyperparameters on their performance. Then I implemented an Echo State Network for predicting the position of the person based on the values received from four capacitive sensors. I optimized the hyperparameters of the implemented network using Hyperopt which is a python library designed for hyperparameter optimization. It automates the process of finding the best hyperparameters for machine learning models, making it more efficient than manual tuning. I compared the model with an existing temporal convolutional network (TCN) in terms of computational costs and accuracy. Then I integrated both models together and again compared them in terms of these criteria. The result shows that however, TCN outperforms the basic ESN, both if we integrate the two networks together, by adding a little complexity we can achieve higher accuracy.